The software company that created FT Mercati Go to ftmercati.com

Case study

Market reports in five languages, written with AI

The engine that makes FT Mercati's PDF reports: charts and tables from the databases, the model's text on the data, a guardrail that stops what is wrong.

In production Periodic market reports for FT Mercati's subscribers, in production since 2026 Updated on

  • 13types of recurring report: eight weekly and five monthly, on metals, exchange rates, freight rates, plastics, oil and steel
  • 5languages for every report: written in Italian, translated into the other four
  • 14FT Mercati databases from which numbers, tables and charts are taken
  • 2model providers, with automatic switchover to the second on a technical error
  • 7steps in the news workflow, from preparing the query to the final check

In brief

FT Mercati's weekly and monthly reports, in PDF and in five languages, are made by an engine built by FT Support: charts, tables and numbers come from the databases; the model writes the commentary on the data it receives, in Italian, and translates it. Two model providers with automatic switchover on failure, and a guardrail that, if the text is missing, is a refusal or is in the wrong language, puts in a placeholder instead of sending it.

The starting point

Every week and every month FT Mercati sends its subscribers reports on the markets: base and precious metals, traders’ positions, exchange rates, brass, scrap, steel products, freight rates, plastics, oil, steel. In five languages, with up-to-date charts and a commentary consistent with the numbers. Written by hand, they would not be sustainable; written “from memory” by an artificial intelligence (AI) model, they would not be reliable.

What we built

A modular system: a module for each type of report, the repositories on the databases, chart generation, PDF layout, scheduling. The sequence is always the same:

  1. the numbers are read from FT Mercati’s databases and prepared by the code: closing prices, changes, averages, stocks, curves, indicators;
  2. the charts are drawn from those numbers;
  3. the model receives the data already calculated and writes the commentary, in Italian;
  4. the commentary is translated into the other four languages;
  5. the PDF is laid out and placed in the folder from which the sending module sends it to each subscriber in their own language, with a log of what is sent and caps per run.

Where the AI works

  • The commentary. Summary, technical reading and reading of the fundamentals, written on the data received and not on the model’s general knowledge; the prompts for each report can be edited from the interface, without touching the code.
  • The news. A workflow in seven steps: preparing the query, web search, selection of the sources, extraction of the sources’ content, writing, critique, publication with a guardrail. Only real news, with the figures injected from the database; as a backup, the archive of the FT News newsroom of agents.
  • The translations. Written once in Italian, then translated into the other languages: fewer calls, more consistency between the versions.

Charts and PDFs are deterministic: same data, same drawing.

How we keep the AI under control

  1. The code calculates the numbers. The model does not query the database: it receives the prepared data and comments on that.
  2. Two providers. A main model and a backup one, with automatic switchover only on a technical error; the models can be changed from the configuration, without recompiling.
  3. A guardrail that blocks. Every text goes through a validator: empty, model refusal, wrong language, too short. If it does not pass, a placeholder in the language of the report goes in its place. The comment in the code next to that check puts it this way: “a visible placeholder instead of shipping garbage to clients”.
  4. Prompt and temperature for each report, which the people who use them can change.
  5. Everything automatic, with generation and sending kept separate. In production since the beginning of 2026, the engine produces the files at a fixed time and another module emails them to each subscriber, with a log of what is sent and caps per run; no person reads through the reports before they are sent, and that is why the guardrail blocks instead of letting things through.

Declared limits

  • the language validator checks only the most frequent case, English where Italian is needed, and only on texts that are long enough;
  • the commentary is only as good as the data it receives: a missing figure produces a skipped section or a notice, not an invention.

What it means for your company

Every company has recurring documents that come from data: the production report, the quality report, the monthly sales summary, the letter to customers. There is one rule: the numbers from the ERP (the company’s management software), the text from the model on those numbers, a guardrail that stops what is wrong. We discuss it in the data and integrations service and in the guide on the AI assistant on company documents. If you have a report that costs a day a month today, let’s talk.

Frequently asked questions

FT Report: reports with AI: questions and answers

Does the AI make up the numbers in the reports?

No. Prices, changes, stocks, averages and indicators are calculated by the code on FT Mercati's databases and passed to the model ready-made; the model writes the commentary on those numbers, and the charts are generated from the data, not by the AI. If a figure is missing, the section is skipped or carries a notice that the content is not available: never a made-up number.

What happens if the model does not respond, or responds badly?

If the main provider returns a technical error, the call goes to the second one. If the text produced is empty, is a refusal by the model or is in the wrong language, the guardrail discards it and puts in a placeholder in the language of the report: a subscriber never receives an empty section, nor a text in English where they expect Italian.

Are the news items cited in the reports real?

Yes. The news workflow searches the web for real sources, filters them, extracts their content and rewrites it with the figures taken from the database; if it finds no verified news for the period, the report says so. As a backup it uses FT Mercati's news archive.

Is it a service that FT Support sells?

The reports are an FT Mercati service for its subscribers. FT Support designed and built the engine that produces them, and this is what it can do again for another company: that company's data, its recurring documents, its languages, with the same rules.

Other case studies

Do you have a process you would like to see working better?

Tell us about it in a first conversation, with no obligation: together we work out whether and where AI brings a measurable benefit, and which step to start with.